GHSA-Q87H-C5M7-659J
Vulnerability from github – Published: 2026-11-20 23:48 – Updated: 2026-11-20 23:48Summary
Every sibling <table> adds another level of the parent's indentation, so the size of get_text() output grows quadratically with the number of tables. In the measurements below, a 240 KB HTML document produces about 1 billion characters and uses about 1 GB of memory.
Details
table_end_handler() (model/tag/table_tag.py) writes the rendered table into the parent element with HtmlElement.write_verbatim_text() (model/html_element.py). For a block-level parent, write_verbatim_text() calls canvas.open_block(), which pushes the parent's padding_inline onto the prefix stack. The following close_block() never pops it; only Canvas.close_tag() calls prefix.remove_last_prefix(). So each table leaves one extra padding entry on the stack.
This has been present since 2.0 (commit 84ec720).
>>> get_text('<ul>' + '<table><tr><td>a</td></tr></table>' * 4 + '</ul>')
' a\n \n a\n \n a\n \n a\n '
PoC
import sys, time, resource
from inscriptis import get_text
n = int(sys.argv[1])
html = "<ul>" + "<table></table>" * n + "</ul>"
t = time.time(); out = get_text(html)
print(len(html), f"{time.time()-t:.2f}s", len(out), resource.getrusage(resource.RUSAGE_SELF).ru_maxrss // 2**20, "MiB")
inscriptis 2.7.5, Python 3.11, macOS (2.0.0 behaves the same):
| tables | HTML | time | output chars | peak RSS |
|---|---|---|---|---|
| 1,000 | 15 KB | 0.02 s | 4.0 M | 31 MiB |
| 4,000 | 60 KB | 0.33 s | 64 M | 154 MiB |
| 8,000 | 120 KB | 1.31 s | 256 M | 463 MiB |
| 16,000 | 240 KB | 5.63 s | 1,024 M | 991 MiB |
When the input doubles, output size and memory grow about 4 times.
Impact
Any application that converts untrusted HTML with inscriptis can be made to exhaust memory with a small document. This includes crawlers, RAG and LLM ingestion pipelines, and inscriptis-api. No authentication and no user interaction are needed.
Fix
self.canvas.write(self, text, whitespace=WhiteSpace.pre)
if self.display == Display.block:
+ self.canvas.current_block.prefix.remove_last_prefix()
self.canvas.close_block(self)
With this change the 240 KB input above produces 288 K characters in 0.14 s with 27 MiB peak RSS. A single table renders exactly as before, and the existing test suite passes.
{
"affected": [
{
"database_specific": {
"last_known_affected_version_range": "\u003c= 2.7.5"
},
"package": {
"ecosystem": "PyPI",
"name": "inscriptis"
},
"ranges": [
{
"events": [
{
"introduced": "2.0.0"
},
{
"fixed": "2.7.6"
}
],
"type": "ECOSYSTEM"
}
]
}
],
"aliases": [
"CVE-2026-132277"
],
"database_specific": {
"cwe_ids": [
"CWE-400",
"CWE-407"
],
"github_reviewed": true,
"github_reviewed_at": "2026-11-20T23:48:51Z",
"nvd_published_at": "2026-11-20T17:18:12Z",
"severity": "MODERATE"
},
"details": "# Summary\n\nEvery sibling `\u003ctable\u003e` adds another level of the parent\u0027s indentation, so the size of `get_text()` output grows quadratically with the number of tables. In the measurements below, a 240 KB HTML document produces about 1 billion characters and uses about 1 GB of memory.\n\n# Details\n\n`table_end_handler()` (`model/tag/table_tag.py`) writes the rendered table into the parent element with `HtmlElement.write_verbatim_text()` (`model/html_element.py`). For a block-level parent, `write_verbatim_text()` calls `canvas.open_block()`, which pushes the parent\u0027s `padding_inline` onto the prefix stack. The following `close_block()` never pops it; only `Canvas.close_tag()` calls `prefix.remove_last_prefix()`. So each table leaves one extra padding entry on the stack.\n\nThis has been present since 2.0 (commit 84ec720).\n\n```python\n\u003e\u003e\u003e get_text(\u0027\u003cul\u003e\u0027 + \u0027\u003ctable\u003e\u003ctr\u003e\u003ctd\u003ea\u003c/td\u003e\u003c/tr\u003e\u003c/table\u003e\u0027 * 4 + \u0027\u003c/ul\u003e\u0027)\n\u0027 a\\n \\n a\\n \\n a\\n \\n a\\n \u0027\n```\n\n# PoC\n\n```python\nimport sys, time, resource\nfrom inscriptis import get_text\n\nn = int(sys.argv[1])\nhtml = \"\u003cul\u003e\" + \"\u003ctable\u003e\u003c/table\u003e\" * n + \"\u003c/ul\u003e\"\nt = time.time(); out = get_text(html)\nprint(len(html), f\"{time.time()-t:.2f}s\", len(out), resource.getrusage(resource.RUSAGE_SELF).ru_maxrss // 2**20, \"MiB\")\n```\n\ninscriptis 2.7.5, Python 3.11, macOS (2.0.0 behaves the same):\n\n| tables | HTML | time | output chars | peak RSS |\n|---|---|---|---|---|\n| 1,000 | 15 KB | 0.02 s | 4.0 M | 31 MiB |\n| 4,000 | 60 KB | 0.33 s | 64 M | 154 MiB |\n| 8,000 | 120 KB | 1.31 s | 256 M | 463 MiB |\n| 16,000 | 240 KB | 5.63 s | 1,024 M | 991 MiB |\n\nWhen the input doubles, output size and memory grow about 4 times.\n\n# Impact\n\nAny application that converts untrusted HTML with inscriptis can be made to exhaust memory with a small document. This includes crawlers, RAG and LLM ingestion pipelines, and `inscriptis-api`. No authentication and no user interaction are needed.\n\n# Fix\n\n```diff\n self.canvas.write(self, text, whitespace=WhiteSpace.pre)\n\n if self.display == Display.block:\n+ self.canvas.current_block.prefix.remove_last_prefix()\n self.canvas.close_block(self)\n```\n\nWith this change the 240 KB input above produces 288 K characters in 0.14 s with 27 MiB peak RSS. A single table renders exactly as before, and the existing test suite passes.",
"id": "GHSA-q87h-c5m7-659j",
"modified": "2026-11-20T23:48:51Z",
"published": "2026-11-20T23:48:51Z",
"references": [
{
"type": "WEB",
"url": "https://github.com/weblyzard/inscriptis/security/advisories/GHSA-q87h-c5m7-659j"
},
{
"type": "ADVISORY",
"url": "https://nvd.nist.gov/vuln/detail/CVE-2026-132277"
},
{
"type": "WEB",
"url": "https://github.com/weblyzard/inscriptis/commit/dc122894cd671a0227f8fdad60464acddaf40fc8"
},
{
"type": "PACKAGE",
"url": "https://github.com/weblyzard/inscriptis"
},
{
"type": "WEB",
"url": "https://github.com/weblyzard/inscriptis/releases/tag/2.7.6"
}
],
"schema_version": "1.4.0",
"severity": [
{
"score": "CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:L",
"type": "CVSS_V3"
}
],
"summary": "Inscriptis: Quadratic memory growth from sibling \u003ctable\u003e elements in get_text()"
}
Drag nodes to arrange them. Double-click a related vulnerability to open it, or a source to visit its link. The Physics button on the left tunes the layout and the simulation; the legend at the bottom left hides or shows nodes by kind and by sighting type.
Sightings
| Author | Source | Type | Date | Other |
|---|
Nomenclature
- Seen: The vulnerability was mentioned, discussed, or observed by the user.
- Confirmed: The vulnerability has been validated from an analyst's perspective.
- Published Proof of Concept: A public proof of concept is available for this vulnerability.
- Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
- Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
- Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
- Not confirmed: The user expressed doubt about the validity of the vulnerability.
- Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.
Vulnerability timeline
Sightings and recorded dates, grouped by day · UTC
How durations are calculated
Following the VTEM framework (Black Hat USA 2015), time to patch starts when the vendor or a coordinator was informed, total time to patch at discovery, and disclosure to patch at public disclosure; vendor response time runs from the report to its acknowledgement. Time to exploit and time unpatched start at the earliest discovery or report, falling back to public disclosure. Unpatched time ends at the first recorded patch, or the calculation time when no patch is recorded. Exploitation to patch shows the interval between those two events; a published proof of concept is a public exploit, not exploitation.
Each result shows its source dates. A negative interval means the end event came first; when date-only inputs leave the order open, the result says so. Date-only inputs are approximate. The earliest evidence wins: KEV report dates and sightings are labelled as such. Records describing another vulnerability are listed but never calculated from, and a PoC alone does not establish exploitation. Missing patch evidence does not prove that no patch exists. Filters below change the event list only.
Loading timeline…
The approach is described in our paper Mapping CVEs to MITRE ATT&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion.
Browse all ATT&CK techniques and the vulnerabilities related to each.
Related by attack behaviour
Vulnerabilities whose description is nearest to this one in the vector space of the CIRCL/vulnerability-attack-technique-biencoder model. This is a similarity search over the bi-encoder space (plain cosine), not a classification, and it has no measured accuracy.